How detection works
Catching an ectopic beat on a single-lead chest strap is a trade-off between missing real events and inventing fake ones. RytmoECG does not try to win that with one clever detector. It runs three independent passes, each tuned for a different moment: while you record, when you open one event, and over the whole recording afterwards.
The amber marker is one premature beat. Whether it becomes an event depends on which pass is looking, and that is the whole story of this page.
The three passes
Same algorithms, three contexts. The earlier the pass, the more conservative it has to be; a live alarm you cannot un-see costs more than a finding you chose to open.
Runs on the streaming signal as each beat arrives. Immediate, but gated hard so a noisy stretch cannot spam events.
A second opinion you trigger by opening one strip. Re-derives a local baseline and judges that window in isolation.
Sweeps an entire recording after the fact by interval criteria, surfaces beats the live pass merged away, and adds them.
Pass 1
As the Polar H10 streams 130 Hz lead I, Pan-Tompkins finds each R-peak, a morphology classifier shapes every beat, and a stack of detectors runs against the rolling beat history. Anything that fires goes through the same capture path, so it inherits every guard before it ever becomes an event.
Wide-complex capture, the rhythm-pattern detector (bigeminy, trigeminy, couplet, triplet, NSVT, idioventricular, PVC-with-compensatory-pause, PAC), long-pause (RR > 2.0 s), the rules + CoreML irregular ensemble, and the isolated single-beat ectopy detector below.
Movement gate (accelerometer std-dev), hard SQI floor (mean SQI ≥ 0.50 over thirty beats), baseline-stability gate (DC drift, spread, peak excursion), and a per-kind cooldown. A run of events chains into one clip instead of dozens.
Live is the only pass that interrupts you. A false "maybe PVC" mid-session erodes trust faster than a quiet miss you can still recover in pass two or three. So live errs toward silence when signal or context is ambiguous.
The bar each pass must clear to fire
Inside pass 1
A single premature beat with a compensatory pause is the classic "one PVC" most people picture. The hard part on one lead is not finding it; it is not drowning in false positives from respiratory sinus arrhythmia, where a normal beat can land 20 % early followed by a longer one purely from breathing. This detector catches the real thing while staying quiet on the impostor.
Prematurity is measured against the median of the six beats the early beat interrupted, not a session-wide average that drifts as your heart rate changes. The beat is judged against the rhythm it actually broke.
At most one of those six beats may deviate more than 25 % from the local baseline. If the surrounding rhythm is already irregular, a single early-plus-pause is just normal variation, and belongs to the dedicated irregular and bigeminy paths, not here.
The early interval plus the pause must sum to roughly twice the baseline; the ventricular signature, not the smoother give-and-take of breathing. The same window the per-event and full-session passes use.
The candidate beat needs SQI ≥ 0.70 so a noise spike is not read as an early beat, an eight-second cooldown stops a borderline stretch from repeating, and chaining folds a burst into one reviewable clip.
The interval signature
Plot successive RR intervals as bars and a real ectopic beat is obvious: one bar drops well below baseline (the early beat), the next jumps well above it (the pause), and the two together sum to roughly twice a normal interval. That sum, not the QRS shape, is what every pass keys on.
A double-counted R-peak makes two short bars, never a short-then-long pair, so it fails the sum test. A missed beat makes one long bar with no preceding short one. The compensatory signature is specific on purpose.
Bar height = interval length. Baseline 850 ms.
Why an interval rule alone floods
In a high-variability heart (RMSSD over 80 ms) a plain "20 % early plus a longer beat" rule fires constantly on respiratory sinus arrhythmia, where RR rises and falls smoothly with each breath. The regular-background gate is what tells the two apart: a smooth swing has no single sharp outlier; a PVC does.
RR rises and falls smoothly; no single outlier.
Regular background, one sharp early beat, one pause.
The honest cost: in very high HRV the gate will sometimes hold back a borderline narrow beat, and occasionally let one through. That is deliberate. On a single noisy lead you cannot separate a narrow PVC from respiratory variation by timing alone; wide-complex beats are caught by their shape on a separate path, and anything missed here is recoverable in passes two and three.
Pass 2
Open any event and tap Advanced Analysis. RytmoECG re-runs the full pipeline against that one thirty-to-sixty-second strip in isolation, with a baseline derived from the window itself. Because it is not the live, streaming classifier, it can flag a beat the live pass merged into the dominant sinus template.
It counts ectopy by timing: any beat 20 % or more early followed by a compensatory pause is ectopic regardless of how the morphology classifier labelled its QRS. This surfaces the PVC the live chain reported as "wide-complex: 0".
Likely arrhythmia, ambiguous, likely normal, or poor signal; with the supporting markers and every measured parameter listed, flag colours and all. It reports even a single ectopic beat, where the live pass deliberately would not.
It never fires an alert; you choose to run it on a strip you are already looking at. That is why it can afford a lower bar than live: the cost of a borderline finding is a line of text, not an interruption.
Pass 3
A multi-hour recording can hide ectopy the live pass merged away beat by beat. From a recorded session you can trigger a full re-scan: RytmoECG reads the stored ECG and sweeps every beat by the same interval criteria as the second opinion, then adds each hit as a reviewable event.
Re-runs Pan-Tompkins and a fresh classifier over the recording (capped so a 24 h Holter cannot pull millions of samples at once), independent of whatever the live session's template state happened to be.
Each ectopic beat becomes an event carrying its own strip window, so you can see what came before and after. The exact pass that catches the beat the live morphology gate merged into sinus.
Hits within a second or so of an event already on the session are skipped, so re-scanning does not pile duplicates on top of what live already captured.
Where to find it: open a recorded session, then run the re-scan from its detail view. It reports how many new events it added.
When the passes disagree
The most common question: "I can see the early beat, Advanced Analysis flags it, so why was there no live event?" Usually the live classifier merged that beat into the dominant sinus template, so the shape-based path saw nothing wide, and the live baseline at that instant was drifting, so the timing looked less premature than it does against a clean window.
The classifier merged it into sinus (wide-complex: 0) and the drifting baseline softened the prematurity. The isolated-ectopy detector closes most of this gap in real time, but live still errs quiet when context is ambiguous.
Judged in isolation against a clean local baseline, by timing alone. The early-plus-pause sum clears the bar, so the strip reads "ambiguous" or "likely arrhythmia" with the markers listed.
The same interval criteria over the whole recording surface it as a reviewable event, deduped against anything live already caught. The catch-all safety net.
The principle: layers, not one verdict. The same interval criteria run in passes two and three, so a finding is reproducible, not a one-off. Nothing leaves the device; all three passes run on your iPhone.
Get it
iPhone, iOS 17 or later. Requires a Polar H10 chest strap.
App Store; coming soonRytmoECG is a wellness app; not a medical device. It does not diagnose, treat, or screen for any condition. Every detector phrases its result as a possibility, never a diagnosis. If you have chest pain, shortness of breath, or persistent palpitations, contact emergency services or your physician.